Inference device, convolution calculation execution method and program

TL;DR: In this paper, an inference device is provided with a quantizer unit, a convolution calculation unit and an input data conversion unit, which, in each layer, shares a circuit for executing convolution calculations using quantized input data.

Abstract: An inference device is provided which, in each layer, shares a circuit for executing convolution calculations using quantized input data. This inference device is provided with a quantizer unit, a convolution calculation unit and an input data conversion unit. The quantizer unit quantizes the results of convolution calculations in a convolution neural network that uses input data and weights. The convolution calculation unit executes convolution calculations using the quantized calculation results as input data. The input data conversion unit converts first-layer input data such that the convolution calculation unit can similarly process the first-layer input data and the input data quantized by the quantizer unit.